1 // Auto-generated file. Do not edit!
2 // Template: src/f16-igemm/avx2-broadcast.c.in
3 // Generator: tools/xngen
4 //
5 // Copyright 2022 Google LLC
6 //
7 // This source code is licensed under the BSD-style license found in the
8 // LICENSE file in the root directory of this source tree.
9
10 #include <assert.h>
11
12 #include <immintrin.h>
13
14 #include <xnnpack/igemm.h>
15 #include <xnnpack/intrinsics-polyfill.h>
16
17
xnn_f16_igemm_minmax_ukernel_7x8__avx2_broadcast(size_t mr,size_t nc,size_t kc,size_t ks,const void ** restrict a,const void * restrict w,void * restrict c,size_t cm_stride,size_t cn_stride,size_t a_offset,const void * zero,const union xnn_f16_scaleminmax_params params[restrict XNN_MIN_ELEMENTS (1)])18 void xnn_f16_igemm_minmax_ukernel_7x8__avx2_broadcast(
19 size_t mr,
20 size_t nc,
21 size_t kc,
22 size_t ks,
23 const void**restrict a,
24 const void*restrict w,
25 void*restrict c,
26 size_t cm_stride,
27 size_t cn_stride,
28 size_t a_offset,
29 const void* zero,
30 const union xnn_f16_scaleminmax_params params[restrict XNN_MIN_ELEMENTS(1)])
31 {
32 assert(mr != 0);
33 assert(mr <= 7);
34 assert(nc != 0);
35 assert(kc != 0);
36 assert(kc % sizeof(uint16_t) == 0);
37 assert(ks != 0);
38 assert(ks % (7 * sizeof(void*)) == 0);
39 assert(a_offset % sizeof(uint16_t) == 0);
40 assert(a != NULL);
41 assert(w != NULL);
42 assert(c != NULL);
43
44 uint16_t* c0 = c;
45 uint16_t* c1 = (uint16_t*) ((uintptr_t) c0 + cm_stride);
46 if XNN_UNPREDICTABLE(mr < 2) {
47 c1 = c0;
48 }
49 uint16_t* c2 = (uint16_t*) ((uintptr_t) c1 + cm_stride);
50 if XNN_UNPREDICTABLE(mr <= 2) {
51 c2 = c1;
52 }
53 uint16_t* c3 = (uint16_t*) ((uintptr_t) c2 + cm_stride);
54 if XNN_UNPREDICTABLE(mr < 4) {
55 c3 = c2;
56 }
57 uint16_t* c4 = (uint16_t*) ((uintptr_t) c3 + cm_stride);
58 if XNN_UNPREDICTABLE(mr <= 4) {
59 c4 = c3;
60 }
61 uint16_t* c5 = (uint16_t*) ((uintptr_t) c4 + cm_stride);
62 if XNN_UNPREDICTABLE(mr < 6) {
63 c5 = c4;
64 }
65 uint16_t* c6 = (uint16_t*) ((uintptr_t) c5 + cm_stride);
66 if XNN_UNPREDICTABLE(mr <= 6) {
67 c6 = c5;
68 }
69
70 do {
71 __m256 vacc0x01234567 = _mm256_cvtph_ps(_mm_load_si128((const __m128i*) w));
72 __m256 vacc1x01234567 = vacc0x01234567;
73 __m256 vacc2x01234567 = vacc0x01234567;
74 __m256 vacc3x01234567 = vacc0x01234567;
75 __m256 vacc4x01234567 = vacc0x01234567;
76 __m256 vacc5x01234567 = vacc0x01234567;
77 __m256 vacc6x01234567 = vacc0x01234567;
78 w = (const uint16_t*) w + 8;
79
80 size_t p = ks;
81 do {
82 const uint16_t* restrict a0 = (const uint16_t*) a[0];
83 assert(a0 != NULL);
84 if XNN_UNPREDICTABLE(a0 != zero) {
85 a0 = (const uint16_t*) ((uintptr_t) a0 + a_offset);
86 }
87 const uint16_t* restrict a1 = (const uint16_t*) a[1];
88 assert(a1 != NULL);
89 if XNN_UNPREDICTABLE(a1 != zero) {
90 a1 = (const uint16_t*) ((uintptr_t) a1 + a_offset);
91 }
92 const uint16_t* restrict a2 = (const uint16_t*) a[2];
93 assert(a2 != NULL);
94 if XNN_UNPREDICTABLE(a2 != zero) {
95 a2 = (const uint16_t*) ((uintptr_t) a2 + a_offset);
96 }
97 const uint16_t* restrict a3 = (const uint16_t*) a[3];
98 assert(a3 != NULL);
99 if XNN_UNPREDICTABLE(a3 != zero) {
100 a3 = (const uint16_t*) ((uintptr_t) a3 + a_offset);
101 }
102 const uint16_t* restrict a4 = (const uint16_t*) a[4];
103 assert(a4 != NULL);
104 if XNN_UNPREDICTABLE(a4 != zero) {
105 a4 = (const uint16_t*) ((uintptr_t) a4 + a_offset);
106 }
107 const uint16_t* restrict a5 = (const uint16_t*) a[5];
108 assert(a5 != NULL);
109 if XNN_UNPREDICTABLE(a5 != zero) {
110 a5 = (const uint16_t*) ((uintptr_t) a5 + a_offset);
111 }
112 const uint16_t* restrict a6 = (const uint16_t*) a[6];
113 assert(a6 != NULL);
114 if XNN_UNPREDICTABLE(a6 != zero) {
115 a6 = (const uint16_t*) ((uintptr_t) a6 + a_offset);
116 }
117 a += 7;
118
119 size_t k = kc;
120 do {
121 const __m256 vb01234567 = _mm256_cvtph_ps(_mm_load_si128((const __m128i*) w));
122 w = (const uint16_t*) w + 8;
123
124 const __m256 va0 = _mm256_cvtph_ps(_mm_set1_epi16((short) *a0));
125 a0 += 1;
126 const __m256 va1 = _mm256_cvtph_ps(_mm_set1_epi16((short) *a1));
127 a1 += 1;
128 const __m256 va2 = _mm256_cvtph_ps(_mm_set1_epi16((short) *a2));
129 a2 += 1;
130 const __m256 va3 = _mm256_cvtph_ps(_mm_set1_epi16((short) *a3));
131 a3 += 1;
132 const __m256 va4 = _mm256_cvtph_ps(_mm_set1_epi16((short) *a4));
133 a4 += 1;
134 const __m256 va5 = _mm256_cvtph_ps(_mm_set1_epi16((short) *a5));
135 a5 += 1;
136 const __m256 va6 = _mm256_cvtph_ps(_mm_set1_epi16((short) *a6));
137 a6 += 1;
138
139 vacc0x01234567 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(va0, vb01234567, vacc0x01234567), _MM_FROUND_NO_EXC));
140 vacc1x01234567 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(va1, vb01234567, vacc1x01234567), _MM_FROUND_NO_EXC));
141 vacc2x01234567 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(va2, vb01234567, vacc2x01234567), _MM_FROUND_NO_EXC));
142 vacc3x01234567 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(va3, vb01234567, vacc3x01234567), _MM_FROUND_NO_EXC));
143 vacc4x01234567 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(va4, vb01234567, vacc4x01234567), _MM_FROUND_NO_EXC));
144 vacc5x01234567 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(va5, vb01234567, vacc5x01234567), _MM_FROUND_NO_EXC));
145 vacc6x01234567 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(va6, vb01234567, vacc6x01234567), _MM_FROUND_NO_EXC));
146
147 k -= sizeof(uint16_t);
148 } while (k != 0);
149 p -= 7 * sizeof(void*);
150 } while (p != 0);
151
152 const __m256 vscale = _mm256_load_ps(params->avx.scale);
153 vacc0x01234567 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_mul_ps(vacc0x01234567, vscale), _MM_FROUND_NO_EXC));
154 vacc1x01234567 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_mul_ps(vacc1x01234567, vscale), _MM_FROUND_NO_EXC));
155 vacc2x01234567 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_mul_ps(vacc2x01234567, vscale), _MM_FROUND_NO_EXC));
156 vacc3x01234567 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_mul_ps(vacc3x01234567, vscale), _MM_FROUND_NO_EXC));
157 vacc4x01234567 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_mul_ps(vacc4x01234567, vscale), _MM_FROUND_NO_EXC));
158 vacc5x01234567 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_mul_ps(vacc5x01234567, vscale), _MM_FROUND_NO_EXC));
159 vacc6x01234567 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_mul_ps(vacc6x01234567, vscale), _MM_FROUND_NO_EXC));
160
161 const __m256 vmin = _mm256_load_ps(params->avx.min);
162 vacc0x01234567 = _mm256_max_ps(vacc0x01234567, vmin);
163 vacc1x01234567 = _mm256_max_ps(vacc1x01234567, vmin);
164 vacc2x01234567 = _mm256_max_ps(vacc2x01234567, vmin);
165 vacc3x01234567 = _mm256_max_ps(vacc3x01234567, vmin);
166 vacc4x01234567 = _mm256_max_ps(vacc4x01234567, vmin);
167 vacc5x01234567 = _mm256_max_ps(vacc5x01234567, vmin);
168 vacc6x01234567 = _mm256_max_ps(vacc6x01234567, vmin);
169
170 const __m256 vmax = _mm256_load_ps(params->avx.max);
171 vacc0x01234567 = _mm256_min_ps(vacc0x01234567, vmax);
172 vacc1x01234567 = _mm256_min_ps(vacc1x01234567, vmax);
173 vacc2x01234567 = _mm256_min_ps(vacc2x01234567, vmax);
174 vacc3x01234567 = _mm256_min_ps(vacc3x01234567, vmax);
175 vacc4x01234567 = _mm256_min_ps(vacc4x01234567, vmax);
176 vacc5x01234567 = _mm256_min_ps(vacc5x01234567, vmax);
177 vacc6x01234567 = _mm256_min_ps(vacc6x01234567, vmax);
178
179 if XNN_LIKELY(nc >= 8) {
180 _mm_storeu_si128((__m128i*) c6, _mm256_cvtps_ph(vacc6x01234567, _MM_FROUND_NO_EXC));
181 c6 = (uint16_t*) ((uintptr_t) c6 + cn_stride);
182 _mm_storeu_si128((__m128i*) c5, _mm256_cvtps_ph(vacc5x01234567, _MM_FROUND_NO_EXC));
183 c5 = (uint16_t*) ((uintptr_t) c5 + cn_stride);
184 _mm_storeu_si128((__m128i*) c4, _mm256_cvtps_ph(vacc4x01234567, _MM_FROUND_NO_EXC));
185 c4 = (uint16_t*) ((uintptr_t) c4 + cn_stride);
186 _mm_storeu_si128((__m128i*) c3, _mm256_cvtps_ph(vacc3x01234567, _MM_FROUND_NO_EXC));
187 c3 = (uint16_t*) ((uintptr_t) c3 + cn_stride);
188 _mm_storeu_si128((__m128i*) c2, _mm256_cvtps_ph(vacc2x01234567, _MM_FROUND_NO_EXC));
189 c2 = (uint16_t*) ((uintptr_t) c2 + cn_stride);
190 _mm_storeu_si128((__m128i*) c1, _mm256_cvtps_ph(vacc1x01234567, _MM_FROUND_NO_EXC));
191 c1 = (uint16_t*) ((uintptr_t) c1 + cn_stride);
192 _mm_storeu_si128((__m128i*) c0, _mm256_cvtps_ph(vacc0x01234567, _MM_FROUND_NO_EXC));
193 c0 = (uint16_t*) ((uintptr_t) c0 + cn_stride);
194
195 a = (const void**restrict) ((uintptr_t) a - ks);
196 nc -= 8;
197 } else {
198 __m128i vh6x01234567 = _mm256_cvtps_ph(vacc6x01234567, _MM_FROUND_NO_EXC);
199 __m128i vh5x01234567 = _mm256_cvtps_ph(vacc5x01234567, _MM_FROUND_NO_EXC);
200 __m128i vh4x01234567 = _mm256_cvtps_ph(vacc4x01234567, _MM_FROUND_NO_EXC);
201 __m128i vh3x01234567 = _mm256_cvtps_ph(vacc3x01234567, _MM_FROUND_NO_EXC);
202 __m128i vh2x01234567 = _mm256_cvtps_ph(vacc2x01234567, _MM_FROUND_NO_EXC);
203 __m128i vh1x01234567 = _mm256_cvtps_ph(vacc1x01234567, _MM_FROUND_NO_EXC);
204 __m128i vh0x01234567 = _mm256_cvtps_ph(vacc0x01234567, _MM_FROUND_NO_EXC);
205 if (nc & 4) {
206 _mm_storel_epi64((__m128i*) c6, vh6x01234567);
207 _mm_storel_epi64((__m128i*) c5, vh5x01234567);
208 _mm_storel_epi64((__m128i*) c4, vh4x01234567);
209 _mm_storel_epi64((__m128i*) c3, vh3x01234567);
210 _mm_storel_epi64((__m128i*) c2, vh2x01234567);
211 _mm_storel_epi64((__m128i*) c1, vh1x01234567);
212 _mm_storel_epi64((__m128i*) c0, vh0x01234567);
213
214 vh6x01234567 = _mm_unpackhi_epi64(vh6x01234567, vh6x01234567);
215 vh5x01234567 = _mm_unpackhi_epi64(vh5x01234567, vh5x01234567);
216 vh4x01234567 = _mm_unpackhi_epi64(vh4x01234567, vh4x01234567);
217 vh3x01234567 = _mm_unpackhi_epi64(vh3x01234567, vh3x01234567);
218 vh2x01234567 = _mm_unpackhi_epi64(vh2x01234567, vh2x01234567);
219 vh1x01234567 = _mm_unpackhi_epi64(vh1x01234567, vh1x01234567);
220 vh0x01234567 = _mm_unpackhi_epi64(vh0x01234567, vh0x01234567);
221
222 c6 += 4;
223 c5 += 4;
224 c4 += 4;
225 c3 += 4;
226 c2 += 4;
227 c1 += 4;
228 c0 += 4;
229 }
230 if (nc & 2) {
231 _mm_storeu_si32(c6, vh6x01234567);
232 _mm_storeu_si32(c5, vh5x01234567);
233 _mm_storeu_si32(c4, vh4x01234567);
234 _mm_storeu_si32(c3, vh3x01234567);
235 _mm_storeu_si32(c2, vh2x01234567);
236 _mm_storeu_si32(c1, vh1x01234567);
237 _mm_storeu_si32(c0, vh0x01234567);
238
239 vh6x01234567 = _mm_srli_epi64(vh6x01234567, 32);
240 vh5x01234567 = _mm_srli_epi64(vh5x01234567, 32);
241 vh4x01234567 = _mm_srli_epi64(vh4x01234567, 32);
242 vh3x01234567 = _mm_srli_epi64(vh3x01234567, 32);
243 vh2x01234567 = _mm_srli_epi64(vh2x01234567, 32);
244 vh1x01234567 = _mm_srli_epi64(vh1x01234567, 32);
245 vh0x01234567 = _mm_srli_epi64(vh0x01234567, 32);
246
247 c6 += 2;
248 c5 += 2;
249 c4 += 2;
250 c3 += 2;
251 c2 += 2;
252 c1 += 2;
253 c0 += 2;
254 }
255 if (nc & 1) {
256 *c6 = _mm_extract_epi16(vh6x01234567, 0);
257 *c5 = _mm_extract_epi16(vh5x01234567, 0);
258 *c4 = _mm_extract_epi16(vh4x01234567, 0);
259 *c3 = _mm_extract_epi16(vh3x01234567, 0);
260 *c2 = _mm_extract_epi16(vh2x01234567, 0);
261 *c1 = _mm_extract_epi16(vh1x01234567, 0);
262 *c0 = _mm_extract_epi16(vh0x01234567, 0);
263 }
264
265 nc = 0;
266 }
267 } while (nc != 0);
268 }
269